Hundred percent system time

vmstat(8) was showing 100% system time on an "idle" system. The problem is that irq_enter() was being called twice for each clock interrupt; once in handle_IRQ_event() and a second time in smp_do_timer():

	do_IRQ()
	|
	+-> handle_IRQ_event()
		|
		irq_enter()				// local_irq_count = 1
			|
			+-> timer_interrupt()
				|
				+-> smp_do_timer()
					|
					irq_enter	// local_irq_count = 2

As a result update_process_times() was attributing every clock tick as occurring during an interrupt and therefore bumping up the per_cpu_system counter.

I could see no other callers of smp_do_timer() other than timer_interrupt(), and this will always be called via handle_IRQ_event. So I believe that the irq_enter/irq_exit in smp_do_timer() is not necessary and incorrect. Is this a true statement? If so, this patch will remove the irq_enter/irq_exit from smp_do_timer.

I've run this patch on a 4 cpu lion in SMP mode and UP mode (CONFIG_SMP turned off as opposed to the nosmp boot option) and have not seen any ill effects and the reported system time seems correct too.

Pat

Patrick O'Rourke
[email protected]

--- linux-test7-plain/arch/ia64/kernel/smp.c	Thu Sep  7 12:26:32 2000
+++ linux-test7-patched/arch/ia64/kernel/smp.c	Thu Sep  7 12:32:00 2000
@@ -457,10 +457,12 @@
 		 * update_process_times() expects us to have done irq_enter().
 		 * Besides, if we don't timer interrupts ignore the global
 		 * interrupt lock, which is the WrongThing (tm) to do.
+		 *
+		 * [ The irq_enter() has been done in our caller, 	]
+		 * [ handle_IRQ_event, so we don't need to do it here.	]
+		 *
 		 */
-		irq_enter(cpu, 0);
 		update_process_times(user);
-		irq_exit(cpu, 0);
 	}
 }

Pat> vmstat(8) was showing 100% system time on an "idle" system. Pat> The problem is that irq_enter() was being called twice for each Pat> clock interrupt; once in handle_IRQ_event() and a second time Pat> in smp_do_timer():

Yes, you are right. This bug was introduced with test7 and, on the x86 side, has been fixed with test8-preXX. The patch looks fine.

	--david